A Two-Stage Deep Generative Model for Masked Face Synthesis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• The COVID-19 pandemic heightened the need for face recognition systems that can handle masked faces.
• Acquiring diverse masked face datasets for training is challenging due to privacy and logistical constraints.
• Existing methods may require facial landmark detection, limiting their applicability.

Purpose of the Study:

• To develop a practical and effective method for synthesizing realistic masked faces from unmasked images.
• To address the scarcity of masked face data for training robust face recognition models.
• To propose a technique that does not rely on facial landmark detection.

Main Methods:

• A novel approach utilizing a cascade of two convolutional auto-encoders (CAEs) for masked face synthesis.
• The first CAE generates a pose-aligned face with a mask pattern.
• The second CAE extracts a segmentation map for accurate mask fusion using image processing.

Main Results:

• The proposed method successfully synthesizes realistic masked faces for unseen individuals.
• The technique is effective even for challenging inputs, including low-resolution (25x25 pixels) and out-of-plane rotated faces.
• Experiments on GTAV Face and Labeled Faces in the Wild (LFW) databases validate the method's accuracy and speed.

Conclusions:

• The cascaded CAE approach provides a simple yet powerful solution for generating synthetic masked faces.
• This method enhances the practicality and robustness of face recognition systems in real-world scenarios.
• The technique offers a viable alternative for data augmentation in masked face recognition research.